Multimodal Multitask Representation Learning for Metadata Prediction in Pathology